🧠 Food for thought

Implications, context, and why it matters.

• The company says its analog processing units (APUs) reach 120 trillion operations per second per watt, about 100x the energy efficiency of top GPUs 1, and can process up to 750x more tokens (the chunks of text these models read and generate) per second per watt than Nvidia’s highest-end GPUs on large language models 2. The 750x figure comes from internal tests on production silicon validated by customers, with no public MLPerf or other third-party results for independent checks against Nvidia 2.
• Software supports ONNX (Open Neural Network Exchange, an open model format) and NVIDIA TensorRT (an inference optimizer for Nvidia GPUs), plus mainstream frameworks 2.
• The company cites validation by the U.S. Department of Defense, auto original equipment manufacturers (OEMs), and defense partners 2. Honda Motor and Lockheed Martin are strategic investors 2. The silicon is in production, and the company raised $125 million, yet it has not named paying customers or design wins 2.
• Data centers face rising power limits as AI could use 10% of U.S. electricity by decade’s end 1. That pressure prompts teams to look past GPU-heavy inference infrastructure (running trained models to generate outputs).
• System integrators (IT services firms that design and deploy hardware/software stacks) with MLOps teams can offer model porting, optimization, plus benchmarking on non-Nvidia accelerators such as Mythic’s APUs 2.
• Service teams can test whether analog compute or other inference chips cut operating costs plus power use for select edge workloads (on-device or near where data is generated) 2. Early skill in quantization (reducing numerical precision to improve efficiency) and retraining workflows (fine-tuning models after conversion) can set these firms apart 3.